Elon Musk Is Going All In on Nvidia, but Are These Two Chip Stocks Better Buys?

AI inference spending is projected to grow at a 32% CAGR through 2032 to $1.3 trillion, roughly double Bloomberg Intelligence's $658 billion estimate for AI training, creating a potentially larger opportunity for AMD and Broadcom than Nvidia. AMD is positioning lower-cost, memory-heavy GPUs and software improvements for inference, supported by two reported $100 billion multiyear deals with OpenAI and Meta and additional GPU customers including Anthropic and Microsoft. Broadcom expects AI revenue to double to $115 billion next fiscal year and double again to $230 billion in fiscal 2028 as custom-chip programs with Alphabet, OpenAI and Meta ramp.

Analysis

The investable issue is not aggregate inference TAM but who captures the cost-per-token curve. ASIC adoption shifts value from merchant accelerators toward hyperscaler-controlled designs, but only after long design-validation cycles; AVGO’s revenue conversion is therefore more visible once customer tape-outs, packaging capacity, and volume commitments are corroborated in earnings rather than inferred from design-win commentary. The second-order beneficiaries are TSMC (TSM), advanced packaging, and high-bandwidth-memory suppliers, while MRVL is the closest public competitive read-through on custom silicon share.

AMD has the highest operating leverage to a credible non-Nvidia accelerator ramp, but it also carries the widest execution gap: software portability, cluster-level networking, and customer deployment evidence matter more than chip specifications. A modest increase in accelerator revenue can drive material estimate revisions because the base is smaller; conversely, discounting hardware to win inference workloads could create revenue upside without the gross-margin upside the equity requires. Treat reported multiyear customer commitments skeptically until they appear as backlog, shipped units, and cash-backed purchase obligations.

Near term, this is likely a sentiment-positive rotation within AI semis rather than a clean zero-sum displacement of NVDA. Over 1-3 months, hyperscaler capex guidance and AVGO/AMD commentary on production ramps are the catalysts; over 6-18 months, inference mix can pressure NVDA’s unit economics only if customers standardize workloads around in-house ASICs or AMD software. The contrarian risk is that low-latency, frontier-model inference remains memory-, networking-, and software-stack constrained, preserving Nvidia’s premium well longer than headline TAM projections imply.

Key Decisions for Investors

• Initiate a 6-month market-neutral long AVGO / short MRVL pair, sized 1:1 by semiconductor beta, only on confirmation that AVGO maintains AI-custom-silicon growth guidance at its next earnings. Thesis is share concentration in the highest-volume hyperscaler programs; exit if AVGO discloses a program delay or if MRVL shows materially faster custom-silicon revenue growth. Target 10-15% relative return; risk 7-8%.
• Keep AMD on a catalyst watch rather than chase the article-driven move. Enter a 3-6 month long only if management quantifies accelerator backlog, raises segment gross-margin outlook, or discloses sustained production deployments at major cloud customers; absent those data, the principal risk is revenue growth bought through price concessions.
• Maintain core NVDA exposure but fund part of it with a modest 6-month NVDA/AMD relative-value hedge: short AMD against long NVDA after AMD outperforms NVDA by 10%+ without an accompanying upward revision to accelerator revenue or gross-margin estimates. This expresses the view that software and systems integration remain the binding constraint in early inference deployment.
• Add TSM to the AI infrastructure basket on broad semiconductor pullbacks rather than after single-stock AI headlines. Custom ASIC proliferation increases leading-edge wafer and advanced-packaging intensity regardless of whether AVGO, AMD, or NVDA wins; key falsifier is a material reduction in hyperscaler capex guidance or a sustained packaging-capacity oversupply signal.